Inspiration
Recurrent Respiratory Papillomatosis (RRP) is a rare but devastating chronic disease caused primarily by HPV serotypes 6 and 11, leading to recurrent tumor growth in the airway and vocal cords. The standard of treatment involves tumor removal via multiple surgical debridement strategies. Patients, especially children, must often undergo multiple surgeries annually simply to maintain breathing and voice function. Patients may also require adjunct medical therapy to attempt remission of lesion regrowth, such as VEGF inhibitors, targeted immunotherapy, antiviral drugs, and more. While multiple approaches exist, there is a lack of an established standard of management, and no single approach has proven to provide a consistent response. Without adequate control of disease, RRP has the rare but devastating risk of malignant transformation to invasive carcinoma, occurring from 1 to 7 percent of patients. Given the highly unpredictable recurrence patterns and repeated airway surgeries that can devastate voice, breathing, and QoL, there is a high burden for this disease. Yet, there is no standardized, patient-specific framework to guide treatment, leaving clinicians to rely on retrospective intuition. Based on this understanding of RPP and its existing limitations we asked ourselves: "What if every new RRP patient could be compared to patients who already lived through similar disease trajectories?" Our innovation: PAPILOT.
What it does
Papilloma Longitudinal Outcome tracking (PAPILOT) is a predictive AI platform for personalized, longitudinal disease modeling for patients with RRP. The platform creates a “digital twin” of each patient and predicts disease trajectory and treatment response by learning from clinically similar patients profiles. PAPILOT leverages a database of ~1,000 patient profiles to match newly diagnosed patients based on demographics (age), disease biology (HPV serotype), phenotypic presentation, and longitudinal disease course (prior surgeries, anatomic extent). Through profile matching, the system predicts response to surgical and medical therapies and ranks potential treatment options to support personalized data-driven clinical decision making.
How we built it
First, we conducted an extensive literature review on PubMed, to identify case studies published up to the present date, focusing on reports that used diverse therapeutic approaches for the treatment of RRP across various patient demographics. After gathering 5 eligible case studies, we extracted data on patient demographics, disease biology, clinical presentation, anatomic extent of disease, medical and surgical interventions, dosing and timing of treatment, and therapeutic response. We encoded symptoms to HPO phenotypic based using HPO (Human Phenotype Ontology) to compare patients using cosine similarity to identify the most clinically similar cases. It then analyzes these nearest neighbors, aggregates their treatment outcomes using weighted evidence, and generates explainable insights to support therapy understanding and decision guidance.
The system visually shows treatment response patterns and clearly explains why patients were matched by highlighting shared symptoms, missing phenotypes, similarity drivers, and therapy outcome evidence. Behind the interface, the platform integrates a FastAPI backend, a digital twin similarity engine, outcome aggregation logic, ontology-based symptom mapping, and synthetic patient cohort generation, creating a complete, deployable prototype that can eventually connect with real clinical EHR systems to support data-driven treatment insights
Challenges we ran into
Our primary challenge was the limited availability of pre-existing data for a rare disease with minimal publicly-available registries and published research. Developing an effective predictive model typically requires large, well-curated datasets, which were not feasible to obtain within the 48-hour hackathon. We addressed this challenge by maximizing the use of existing literature to synthesize a hybrid cohort that helped us prototype a preliminary model.
Accomplishments that we're proud of
We are proud of being able to create a hybrid cohort with limited published data and build a proof-of-concept predictive model for a rare disease in 48 hours. We were able to work well as a team of students from diverse academic backgrounds.
What we learned
Through yesterday’s patient organization talk, we gained insight into RRP and the complex challenges associated with long-term disease management, including the need for multidisciplinary care and the significant impact on QoL. We also learned about the application of digital twinning in non-medical settings and in chronic diseases such as metabolic and cardiac disorders, and explored how this approach could be adapted to rare diseases.
What's next for Papilot
Next steps for PAPILOT include systematically incorporating real-world patient data, beginning with a single-center pilot study and progressing to multi-center expansion to improve cohort diversity and generalizability. We will validate the model’s performance by evaluating its accuracy in predicting disease progression and treatment outcomes across different patient profiles and therapeutic strategies. As the dataset grows, we will continue to refine the model to generate a more comprehensive, clinically meaningful predictive framework that supports decision-making for clinicians. This includes modeling personalized therapeutic pathways based on patient-specific characteristics, optimizing dosing and timing of therapy, and comparing expected outcomes across available treatment options. Ultimately, our goal is to develop a tool that can assist clinicals in selecting personalized management approaches to improve disease control and patient outcomes for RRP.
Log in or sign up for Devpost to join the conversation.